Ten modules covering the math, the architectures, and the code — from linear algebra through transformers, diffusion models, GANs, and beyond. Each builds on the last.
Three phases, ten weeks. Each phase builds on the last — or skip ahead with an alternative path below.
Each module includes practitioner-oriented notes, a hands-on PyTorch notebook, and AI-generated study materials.
Linear algebra, probability, optimization — the math that recurs in every generative model. Softmax, layer norms, residual connections, KL divergence, backpropagation.
Attention, multi-head attention, and the full transformer block — the backbone of modern AI. The single most important module in the series.
AE → VAE → VQ-VAE: the encoder-decoder paradigm and latent spaces. VAEs power Stable Diffusion's latent space; VQ-VAE connects vision to language.
DDPM, DDIM, classifier-free guidance, and latent diffusion (Stable Diffusion). The dominant approach for image and video generation.
Adversarial training, StyleGAN, and fast high-quality generation. A different way of thinking about generation — still state-of-the-art for faces.
Normalizing flows and flow matching — the theoretically cleanest generative models, with exact likelihood computation. Increasingly important and elegant.
GPT, next-token prediction, RLHF, and scaling laws. The paradigm behind LLMs — if your primary interest is language models, start here after Transformers.
Seq2seq, T5, BART, and cross-attention for conditional generation. The architecture behind translation, summarization, and cross-modal attention.
CLIP, text-to-image (DALL-E, Stable Diffusion), image-to-text (LLaVA), video generation (Sora), and the convergence thesis — all modalities becoming tokens.
Sparse routing, load balancing, Switch Transformer — scaling via conditional computation. How transformers learned to grow vast by only consulting a few specialists at a time.
Four curated sequences for different goals. Each one cherry-picks the modules that matter most for your focus.